QsNetIII an Adaptively Routed Network for High Performance Computing Duncan Roweth, Quadrics Ltd Hot Interconnects August 2008
28/8/2008
Quadrics Ltd
1
Quadrics Background
• Develops interconnect products for the HPC market – HPC Linux systems – AlphaServer SC systems
• Quadrics is owned by the Finmeccanica group • Quadrics was 12 years old in July
28/8/2008
Quadrics Ltd
2
QsNet Networks
• Multi-stage switch network • Components – – – – –
Adapter: Elan Router: Elite Switches, cables Firmware, drivers, libraries Diagnostics, documentation
• HPC specific features – Adaptive routing – Hardware barrier & broadcast
28/8/2008
Quadrics Ltd
3
Virtual Address
Communication Model
Processs
28/8/2008
Quadrics Ltd
4
Quadrics Networks
• Elan1 / Elite1, 1994, Meiko Computing Surface 2 – Source chooses between pre-defined routes
• Elan3 / Elite3, 2000, first Quadrics product, QsNet – First use of packet-by-packet adaptive routing – Crosspoint router, x8
• Elan4 / Elite4, 2004, QsNetII – Reduced latency, increased bandwidth – Increased support for offloading collectives
• Elan5 / Elite5, 2008, QsNetIII – General purpose crosspoint router, increased radix, x32 – Highly programmable adapter
28/8/2008
Quadrics Ltd
5
What is Adaptive Routing ?
• Switch networks typically provide many paths between any two points • In an adaptively routed network routers make packet by packet decisions on the route to use based on – – – –
Queue occupancy Channel usage Error rates and state Class of traffic
28/8/2008
Quadrics Ltd
6
Why is Adaptive Routing Important ?
• Most HPC networks are statically routed – They use pre-determined paths between nodes
• Static routing can work well – – – –
If traffic pattern is known in advance If traffic pattern is persistent If traffic pattern is uniform (i.e. application is load balanced) If there are no errors
• These conditions are not met by real codes on production HPC systems {see LLNL and Sandia results} • Adaptive routing solves these problems – Delivering significantly better aggregate bandwidths and worst case latencies on real systems running real codes
28/8/2008
Quadrics Ltd
7
Benefits of Adaptive Routing
• Bandwidth achieved when 1024 nodes all communicate at the same time • Plots show the distribution of measured bandwidths
System
Interconnect
Min
Max
Average
Atlas
Infiniband
95
762
263
Thunder
QsNetII
248
403
369
Data from Lawrence Livermore National Lab, published at the Sonoma OpenFabrics workshop April 2007
28/8/2008
Quadrics Ltd
8
Benefits of Adaptive Routing
• Classic QsNetII all-to-all bandwidth scaling graph
28/8/2008
Quadrics Ltd
9
Ordering Considerations
• Adaptively routed packets can arrive out of order – Problems for stream devices, e.g. multipath Ethernet
• Message ordering is required in HPC – But within a message we are free to deliver the bulk data in arbitrary order Get it there as fast as possible then tell me that it is done
• QsNet ordering – Packets contain the destination virtual address at which to write the data – Bulk data transfers can arrive out of order and can be replayed – Atomic transactions are sequenced
28/8/2008
Quadrics Ltd
10
Adaptive Routing in QsNetIII • More flexible than QsNetII – Operates over arbitrary sets of links – More opportunities to use the technique – Higher radix switches
• Select a subset of lightly loaded output ports based on: – Destination – Link state, errors etc – Number of pending acks (programmable threshold)
• Programmable algorithm for selecting from this subset: – First free, next free, random
28/8/2008
Quadrics Ltd
11
Adaptive Routing: standard case – All top switches are equivalent, select one – Adaptive routing selects a lightly loaded path
28/8/2008
Quadrics Ltd
12
Implementation of Fat Tree Networks
• Connect M×N-way node switches by N×M-way top switches • In this case M = 16, N = 4
28/8/2008
Quadrics Ltd
13
Adaptive Routing in the Top Switch
• If top switch radix ≤ router radix / 2 – i.e. 16 for Elite5, 2048-way networks
• Router provides multiple top switches – Select which to use based on load
• Example: – Traffic from A to B via routers 210 and 300 is blocked by traffic between 300 and 200. – The router providing 300, 301, 302 and 303 can select a different path
28/8/2008
Quadrics Ltd
14
Adaptive Routing on the Final Hop
• Multiple connections to a node • Switch can select a free path • Reduces end-point contention • Simple case is not optimal • Spreading the connections – Improves fault tolerance – Reduces network contention
• Routing decision is made higher in the network
28/8/2008
Quadrics Ltd
15
Adaptive routing in the presence of errors
•
• •
In a production system with 1000s of links it is not uncommon for a small number to be broken – until the next maintenance slot Adaptive routing minimises the impact Example: – Link between routers 10 and 20 is broken – Router 10 dynamically selects paths via 21,22,23 spreading the load. – Reverse case, avoid sending to 10 via 20. Reset 20’s links or update switches 11,12,13.
28/8/2008
Quadrics Ltd
16
Small Packet Support
• Aim to get as close to line rate as possible with small packets • For example: – Small put – 32 byte packet
• Adapter has multiple packet engines • Adapters support up to 64 outstanding packets per link – Doubles if we use both links
• Switches provide 32 virtual channels per output link • Prioritisation – buffering on input to the router 28/8/2008
Quadrics Ltd
17
Barrier & Broadcast Support
• Switches broadcast over a range of output links • Combine Acks / Nacks • Contiguous in QsNetII • Sparse in QsNetIII • Barrier implementation – Network conditional – Broadcast release
28/8/2008
Quadrics Ltd
18
Elan5 – Device Overview
• 2× –
QsNetIII
links
CX4/ QSNetIII
Link
Link
Packet Engine
Packet Engine
Packet Engine
Packet Engine
Packet Engine
Packet Engine
Packet Engine
16K inst cache 9K data buffers
16K inst cache 9K data buffers
16K inst cache 9K data buffers
16K inst cache 9K data buffers
16K inst cache 9K data buffers
16K inst cache 9K data buffers
16K inst cache 9K data buffers
20Gbit/s/direction after protocol
• PCIe, PCIe2 host interface • Multiple packet engines • 512KB of high bandwidth on chip local memory • SDRAM interface to optional local memory • Buffer manager, object cache • Details in ISC Dresden Paper
28/8/2008
CX4/ QSNetIII
Elan5 Adapter
Fabric x8
Host I/F
Local Memory
Object Cache Tags Buffer Manager
Cmd Launch
Free List
PCIe SERDES
PCIe 16 Lanes
Quadrics Ltd
Bridge
Local Functions
TLB
External cache SDRAM i/f
Ext i/f
16K x 8 x 8 banks = 1MB ECC RAM
PLL
External EEPROM DDRII
Clocks
19
Elite5 – Device Overview
• 64 × 32 crosspoint router – Direct & buffered input from each link – 8K of input buffering per link
• • • • •
32 virtual channels per link Physical layer DDR XAUI (6.25GHz) Adaptive routing Hardware barrier and broadcast Memory mapped stats & error counters accessed out-of-band
28/8/2008
Quadrics Ltd
20
QsNetIII Device Overview
Elan
Elite Semi custom ASIC
Manufacturing partners LSI / TSMC G90 process
500 MHz
312 MHz
High performance BGA package
28/8/2008
672 pin
982 pin
< 17W
< 18W
Quadrics Ltd
21
QsNetIII Implementation
• Node switch chassis – 128 links down to the nodes – 128 links up to the top switches – Backplane connects 2 sets of cards
• Top switches – 256 links down to the node switches – Range of system sizes:
28/8/2008
Ports
Radix
Per Chassis
512
4
64
1024
8
32
2048
16
16
4096
32
8
QsNetIII switch logical design
QsNetIII switch implementation
Quadrics Ltd
22
QsNetIII Network 1024–way
• Fat tree, constructed from 8 × 128-way node switches connected by 128 × 8-way top switches
28/8/2008
Quadrics Ltd
23
QsNetIII Implementation – Cables
• • • •
QSFP connectors throughout Copper cables (e.g. Gore) 1-10m Active copper cables (e.g. Gore), 8-20m Optical cables (e.g. Luxtera), 5-300m – PVDF Plenum rated – LSZH available as an option
• No longer Quadrics proprietary
• Likely usage: – Short copper cables from nodes – Optical cables between switches
28/8/2008
Quadrics Ltd
24
QsNetIII Fault Tolerance
• All of the QsNetII Features – – – – –
CRCs on every packet Automatic retransmission Redundant routes Adaptive routing avoids failed links Redundant, hot plugable, PSUs and fans
+ Line rate testing of each link as it comes up – Switches generate CRPAT, CJPAT or PRBS packets – Links are only added to the route tables when they are (a) up, (b) connect to the right place, and (c) can transfer data at full line rate without error.
28/8/2008
Quadrics Ltd
25
QsNetIII Implementation – HP BladeSystem
Elan5 mezzanine adapter 2 QsNet links, PCI-E x8 Gen2 128 MB of memory
28/8/2008
Elite5 switch module Full bandwidth 16 links to the blades (via backplane) 16 links to back of the module
Quadrics Ltd
26
Current Status
• Elite5 silicon in Bristol • Elan5 at TSMC, first parts expected in 3-4 weeks • Switch PCBs, chassis, backplane, controllers are working • First adapter PCBs are ready – PCI-Express x16, HP Blade, ExpressModule (Sun Blade)
• We are porting the QsNetII software • Components at SC08 in Austin • First customer shipment in Q1 of 2009 28/8/2008
Quadrics Ltd
27
Future Work
• QsNetIII hardware – Low cost 32-way switch – 1024-way single chassis switch
• QsNetIII Software – General framework for optimised collectives – Support for “multiport” networks - “fat” nodes have multiple connections to the same rail – Ethernet firmware for the network adapter
28/8/2008
Quadrics Ltd
28
Conclusions
• Adaptive routing underwrites the scalability of HPC systems designed to run a single large application • Adaptive routing has been a feature of QsNet systems since 2000 • QsNetIII offers significant enhancements over both QsNetII and competing products
28/8/2008
Quadrics Ltd
29
Thank you for listening
28/8/2008
Quadrics Ltd
30
Additional Material
28/8/2008
Quadrics Ltd
31
Packet Format
• Packet size of up to 4K made up of 256 byte packet segment and continuations, 8 byte ACK
28/8/2008
Quadrics Ltd
32
Impact of static routing on latency
Data from Thunderbird cluster, Sandia National Lab Big increases in worst case latency with number of nodes
28/8/2008
Quadrics Ltd
33
Impact of static routing on latency
Data from Thunderbird cluster, Sandia National Lab Big variation in worst case latency across a large job
28/8/2008
Quadrics Ltd
34
Software Model – Firmware & Drivers
• Base firmware in the ROMs • Firmware modules loadable with the device driver – Elan, OpenFabrics, 10GE Ethernet, …
• Kernel modules – elan5, elan, rms
• Device dependent library (libelan5) • Device independent library (libelan) • User libraries
28/8/2008
Quadrics Ltd
35
Software Model – Elan Libraries
• Point-to-point message passing • One-sided put/get • Transparent rail striping
28/8/2008
• Optimised collectives • Locks and atomics ops • Global memory allocation
Quadrics Ltd
36
QsNetIII Performance Summary
• Similar latencies to QsNetII – The 1.3 to 2 microsecs of latency is mostly in the host PCI and memory system
• Higher issue rates – Improved link utilisation on small transfers
• Higher bandwidths – 1.5 to 2.25 GB/sec/link depending on host interface
• Bi-directional host interface – 2 x improvement over QsNetII
• Broadcast and barrier in hardware • Continued development of adaptive routing underwrites scaling to high node counts
28/8/2008
Quadrics Ltd
37